AIght_
ToolsCompareLearnUniverseWorkflowsSignalHuman
Take the quiz
← All concepts

Concept

Gradient Descent

Rolling downhill toward a better model.

Moon·Updated Jun 10, 2026
◈ Learn it interactively8 steps · ~6 min →

Where this idea lives

PREREQUISITESGradient DescentLoss FunctionsLoss Functions — The single number a model tries to shrink.BackpropagationBackpropagation — How blame gets assigned through a network.Common misconception: Training searches through possible models until it finds the best one.Common misconception: The algorithm can see where the lowest point is.
prereqsrelatedtoolsmisconceptions
You might think:Training searches through possible models until it finds the best one.The algorithm can see where the lowest point is.

Common misconception

“The algorithm can see where the lowest point is.”

It can't see anything. At each step it knows exactly one fact: the slope of the ground under its feet. It steps downhill, re-measures, steps again. All of deep learning is this near-blind shuffle, repeated trillions of times.

Picture the loss as a landscape. Every possible setting of the model's parameters is a location; the altitude at that location is how wrong the model is there. Somewhere in this terrain are low valleys — parameter settings where the model performs well. Training is the search for low ground.

The catch: for a real model this landscape has billions of dimensions, and there's no map. You can't survey it. The only measurement available is local: standing right here, which direction is downhill? That direction is the gradient, and the algorithm is exactly what you'd guess — take a small step downhill, then ask again.

◉ INTERACTIVE

step 0 · loss 6.14

The ball only ever feels the slope under its feet. Small learning rate: slow but steady. Crank it past ~2 and watch it overshoot the valley entirely.

The one knob that matters

How big should each step be? That's the learning rate, and it's the most consequential setting in training:

  • Too small — convergence takes geological time, and the ball can settle in the first shallow dip it finds.
  • Too large — it overshoots the valley, ricochets between slopes, or diverges entirely. (Crank the slider above past 2 and watch.)

Real training schedules treat the learning rate like a landing: fast early, gentle at the end.

Wrinkles worth knowing

In practice the slope isn't even measured on all the data — that would be too slow. Each step uses a small random batch of examples, making the descent slightly drunk: noisy, wandering, and — usefully — able to stumble out of bad shallow valleys. This is stochastic gradient descent, the workhorse of the field.

And the landscape for deep networks isn't a tidy bowl; it's a billion-dimensional mountain range. The minor miracle the modern era rests on: in very high dimensions, this stumbling walk reliably finds valleys that are good enough. Nobody fully knows why it works this well. It just does, every day, on every model you've used.

What to read next

Downhill on the training data, though, can mean memorizing it. Overfitting and underfitting — the tension that defines what "a good model" even means.

◈ Check yourself

You just read it — now see if it stuck. No score, no penalty for a wrong pick.

Question 1 of 3

Gradient descent improves a model by…

Question 2 of 3

The classic metaphor: gradient descent is like walking downhill in fog. What's the "hill"?

Question 3 of 3

Why does the learning rate (step size) matter in gradient descent?

← Prerequisite

Loss Functions

Next →

Overfitting & Underfitting

by Moon · poet · engineer · third-year CS, Chandigarh University · Updated Jun 10, 2026

← Back to all conceptsBrowse tools →
beginner
Read time6 min read
UpdatedJun 2026
Sources2

Read next

  1. Loss Functions →
  2. Backpropagation →